Institute of Marine Electronic and Intelligent System, Ocean College, Zhejiang University, Zhoushan 316021, China.
The Engineering Research Center of Oceanic Sensing Technology and Equipment, Ministry of Education, Zhoushan 316021, China.
Sensors (Basel). 2022 Aug 18;22(16):6197. doi: 10.3390/s22166197.
Monocular 3D object detection is very challenging in autonomous driving due to the lack of depth information. This paper proposes a one-stage monocular 3D object detection network (MDS Net), which uses the anchor-free method to detect 3D objects in a per-pixel prediction. Firstly, a novel depth-based stratification structure is developed to improve the network's ability of depth prediction, which exploits the mathematical relationship between the size and the depth in the image of an object based on the pinhole model. Secondly, a new angle loss function is developed to further improve both the accuracy of the angle prediction and the convergence speed of training. An optimized Soft-NMS is finally applied in the post-processing stage to adjust the confidence score of the candidate boxes. Experiment results on the KITTI benchmark demonstrate that the proposed MDS-Net outperforms the existing monocular 3D detection methods in both tasks of 3D detection and BEV detection while fulfilling real-time requirements.
由于缺乏深度信息，单目 3D 目标检测在自动驾驶中极具挑战性。本文提出了一种基于无锚点的单目 3D 目标检测网络（MDS Net），该网络采用逐像素预测的方法来检测 3D 目标。首先，开发了一种新颖的基于深度的分层结构，以提高网络的深度预测能力，该结构利用针孔模型中物体图像的大小和深度之间的数学关系。其次，开发了新的角度损失函数，以进一步提高角度预测的准确性和训练的收敛速度。最后，在后处理阶段应用了优化的 Soft-NMS 来调整候选框的置信得分。在 KITTI 基准上的实验结果表明，所提出的 MDS-Net 在 3D 检测和 BEV 检测任务中均优于现有的单目 3D 检测方法，同时满足实时要求。
